Improving the efficiency of interpretability techniques for unstructured Machine Learning problems

Machine Learning (ML) is rapidly changing the world, having an increasingly important role in many areas of everyday life. Entire industries, like healthcare, manufacturing and automotive have been completely revolutionized by the huge impact that research is having in the Artificial Intelligence (AI) field. Data plays an important role in this revolution: the enormous amount of data collected, together with the rise of new technologies to process them, brought to the development of new ML techniques. Nevertheless, many of them still face issues that slow down their adoption. One of the most important is presented by their increasing complexity causing a lack of trust in their decisions and outputs. Several techniques have been developed to increase trust in ML model decisions however they can be computational demanding especially when in presence of unstructured data (images, signals, time series, etc.). This thesis deals with the topic of Explainable AI (XAI) for unstructured data. We test and evaluate new approaches with a focus on the execution time and the quality of the final results. Methods that exploit generative ML models are explored, with a particular attention on Variational Autoencoders (VAE). These kinds of techniques are already being used by the company where this work was carried out, to analyze and evaluate tabular datasets. The aim of the thesis is to develop an interface that allows the extension of their methods to image datasets, without losing their benefits in terms of computational complexity and quality of the results. Different approaches are evaluated to reduce complexity in unstructured data, starting from classic techniques, such as Wavelet transforms, to the use of Convolutional Neural Networks (CNN), finally developing hybrid techniques that exploit both of these methods. The outputs of the interface are then used to feed a Tabular VAE, to produce the final analysis on the initial data. Many limitations were taken into account during the work since this is a first approach to this problem, like the number of channels and the size of the initial images. The results have been analyzed and evaluated in order to find the most promising approach. Both qualitative and quantitative analysis have been conducted, examining the reconstruction error of the inputs and their latent space distribution inside the VAE. We finally present some considerations about the designed techniques, together with possible future developments for the work.